The indexing process required by content-based video databases is correlated to visual characteristics of their content. A preliminary step consists in the partitioning of the video into a sequence of short dynamic scenes, generally characterized by a set of homogeneous features. Each scene may be therefore characterized by the features of one of more representative frames, i.e. still images. In this paper, a method for scene detection of MPEG-1 and MPEG-2 video sequences is reported. The method does not need to decode the streams, because it is based on the analysis of their external characteristics, such as the frame pattern and the sized of I-, P- and B-frames. Changes in above characteristics are used to detect the frames representing each scene, using heuristics and statistical considerations. Since the analysis is based on very simple computation, the algorithm performs very fast. Its computational cost is linear dependent on the number of frames. On the other hand, the method is not well suited for video clip of short length, for which a statistical analysis is not significant. For its low computation cost and accuracy, the method is a suitable tool for preliminary segmentation step of long video sequences.